Stacked Ensemble Machine Learning for Phenological Wheat Yield Prediction from UAV-RGB and multispectral Satellite Data

Highly accurate monitoring of crop growth and yield prediction is essential for farm-level field management and ensuring food security. High resolution satellite and unmanned aerial vehicle (UAV) imaging systems offer timely and cost-effective methods of agroecosystem monitoring under changing climatic conditions. In this context, this study aimed to monitor farm-level wheat phenological growth and yield prediction by integrating UAV-RGB and satellite-based multispectral (Sentinel-2) time series data in a stacked machine learning framework. Six UAV flights were conducted following the wheat crop growth calendar at an altitude of 50 m on a 25-acre experimental site in the Bahawalpur District, Pakistan. The UAV-based Visible Atmospheric Resistance Vegetation Index (VARI) was computed from each Ortho mosaic image presenting a specific stage of crop growth. In parallel, five satellite-based vegetation indices, namely the Normalized Difference Vegetation Index (NDVI), Normalized Difference Red Edge Index (NDRE), modified soil-adjusted vegetation index (MSAVI), Normalized Difference Moisture Index (NDMI), and Red Edge Chlorophyll Index (RECI), were also computed from Sentinel-2. All vegetation indices were used as inputs into three separate stacked machine learning frameworks to predict wheat yield across 22 acres, independently at the vegetative (Vg), reproductive (Rp), and maturity (Mt) stages of crop growth. The results revealed that the ensemble of four base models showed R 2 = 0.70, RMSE = 2.44 at the Vg stage, R 2 = 0.79, RMSE = 2.05 at the Rp stage, and R 2 = 0.75, RMSE = 2.20 at the Mt stage. The GBR meta-learner improved the prediction performance at all stages compared to the ensemble, with the highest R 2 = 0.83 and lowest RMSE = 1.78 achieved at the Mt stage. It was followed by R 2 = 0.79, RMSE = 2.02 and 2.03 at Rp and Vg stages. Furthermore, the SHAP kernel explainer showed that MSAVI, VARI and NDRE contributed significantly to yield predictions. These findings underscore the significance of integrating UAV-RGB and high-resolution multispectral data for accurate and timely wheat yield predictions across 22 plots in a single season. It reflects the effectiveness of proposed stacked machine learning framework supporting stakeholders in making sustainable decisions in limited agricultural settings. • UAV-RGB integrated with multispectral Sentinel-2 indices provided accurate identification of wheat growth stages. • Among the base models, Random Forest achieved highest prediction accuracy in reproductive stage. • Meta learner GBR outperformed base models and ensemble models at maturity and reproductive stage. • Reproductive and Maturity stages are the most suitable periods for accurate yield modeling.